Research map / Probabilistic and scientific modeling / Causal discovery
When Selection Meets Intervention: Additional Complexities in Causal Discovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces an interventional twin graph model and CDIS, a sound FCI-based algorithm, for causal discovery from soft interventions with unknown targets under selection bias.
Problem
Interventional causal discovery ignores selection bias where subjects are pre-selected before interventions; existing augmented DAG and FCI-based paradigms mis-model such data and yield false causal conclusions.
Approach
Models the observed (reality) world and a counterfactual basal world where selection occurs before intervention via an interventional twin graph; derives Markov properties, MAG-based graphical criteria for Markov equivalence, and a sound algorithm (CDIS) that uses pooled observational and interventional data to orient edges and selection structure.
Key results
- On synthetic linear SEM data, CDIS achieves average precision above 0.7 across most configurations, while baselines generally stay below 0.65 (except UT-IGSP in some instances).
- Algorithm is proven sound under oracle CI tests; completeness is only conjectured, supported by a search over about 50,000 (graph, target) pairs with up to 15 vertices with no counterexample found.
- On sciPlex2 gene regulation, the method recovers validated relations such as RELA to RUNX1 and JUNB to MAFF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDIS instead of existing causal discovery baselines | Interventional causal discovery under selection bias · Simulations, most configurations | precision: CDIS exceeds 0.7; baselines generally below 0.65 | Helped |
Surprising findings
- Under selection, an intervention can add dependencies not present in the original DAG, so interventional distributions may not be Markovian to the original graph.
- Directly applying standard orientation rules to skeletons from observational data can falsely orient edges, so orientation must use denser interventional adjacencies.
Limitations
- Completeness of CDIS is conjectured but not proven.
- Assumes causal sufficiency (no latent confounders beyond the model) and faithfulness.
- No graphical representation for the full equivalence class is yet developed; post-intervention selection is not modeled.
Experimental setup
Datasets: sciPlex2 educational random controlled trial dataset (college freshmen incentives and services) synthetic linear SEM data (5,000 samples per setting)
Benchmarks: sciPlex2 educational RCT dataset
Metrics: precision recall F1 score of directed edges edgemark accuracy structural Hamming distance
Baselines: GIES IGSP UT-IGSP CD-NOD JCI-GSP
Builds on: FCI augmented DAG twin network maximal ancestral graph GIES
Closest papers by meaning
- Latent Variable Causal Discovery under Selection Bias (ICML 2025 · similarity 0.88)
- Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.88)
- Bivariate Causal Discovery using Bayesian Model Selection (ICML 2024 · similarity 0.88)
- Identifying Structural Biases from Causal Mechanism Shifts (NeurIPS 2026 · similarity 0.87)
- Continuous Bayesian Model Selection for Multivariate Causal Discovery (ICML 2025 · similarity 0.87)
- Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.